TISR: Twin Image Super-Resolution using Deep Convolutional Neural Networks

Image super-resolution aims to reconstruct the visually pleasing high-resolution (HR) image from the degraded version or low-resolution (LR) ones. Under the remarkable improvement in the field of image and computer vision tasks. Despite its improvement in accuracy and performance, these models used convolutional neural network (CNN) layers side by side to increase the depth of the network, which is not a suitable way of design and it creates the vanishing gradient problems during the training. Furthermore, previous deep CNN methods rely on a single channel to reconstruct the HR output image, but later end layers cannot receive the proper information and work as a dead layer. In this paper, we are used two parallel deep convolutional neural networks with the same size and order of filters, known as Twin Image Super-Resolution using Deep Convolutional Neural Networks (TISR). Additionally, proposed method used two parallel branches for extracting the low, mid, and high-level features simultaneously. For multi-level feature extraction purposes latest Xception block is employed from GoogLeNet architecture. Our method is evaluated on three different benchmark test datasets including SET5, SET14, and BSDS100. Experimental results are demonstrated that our proposed method (TISR) outperforms then the existing state-of-the-art methods.